AI is saying the wrong thing about your brand: how to find it, fix it, and prove it changed

Start with what the answer actually says and which evidence it provides. Your own pages may need correcting, but the source may also sit elsewhere or remain unknown.

The first 48 hours: what to do right now

If you have just seen an AI engine say something false about your business, do these four things before you do anything else.

Capture it properly. Screenshot the answer, note the exact prompt wording, the engine, the date and whether you were logged in. One screenshot is evidence that it happened once, not evidence of a pattern.

Run the same prompt three more ways. Reword it, try it in a different engine, and try it in a fresh session with no history. If it appears once in four, you have noise. If it appears four times in four, you have a problem.

Ask where it came from. In engines that show citations, open the cited pages. The wrong claim is either sitting on one of them, or it is not, and that single distinction determines everything you do next.

Do not email OpenAI. There is no support queue that edits what a model says about your company, and waiting on a reply costs you the weeks you should be spending on the source layer.

What does it mean when AI gets your brand wrong?

It helps to separate three things that get lumped together.

Inaccuracy is a false statement of fact. You are described as having 12 staff when you have 60, or as serving retail when you serve manufacturing.

Staleness is a true statement that has expired. Your old pricing, a client you no longer work with, a service you retired eighteen months ago.

Misattribution is your identity blurring into something else. A competitor’s product credited to you, or your name attached to a different company that shares it.

They feel identical when you read the answer. They have different causes and different fixes, which is why the first real step is diagnosis rather than reaction.

Why does AI say things about my brand that are not on my website?

Because your website is a minority shareholder in your own description.

When an answer engine builds a description of your company, it retrieves from across the web and synthesises. Directories, review platforms, trade press, forums, partner sites, aggregators and databases you have never logged into all feed the same answer your website does.

Exactly how much of it is yours is genuinely contested. Published estimates of the share of AI-cited sources that come from a brand’s own domain range from low single digits to well over forty percent, and the spread is mostly definitional: whether “owned” means your website alone or your website plus every listing and profile you control, and whether the queries measured were branded or unbranded. Anyone quoting you one confident number is quoting one study’s definition.

What is not contested is the direction. Your website is one input among many, it is not treated as the authority on you, and no reading of any of those studies makes it the bulk of what the engine reads.

This is the part that catches people out. The old model was simple: your website was the record, so if the record was wrong you edited the page. The new model is that AI engines read the whole internet and then answer for you. You can rewrite your About page on Monday and still be described using a 2023 directory listing on Friday, because the engine never treated your About page as authoritative in the first place.

There is a second reason, and it is the one most guides skip. Retrieval systems do not return “unknown” when they lack information. They produce the most probable answer given what they have. If the web is thin on your company, the engine will still generate a confident description, built by inference from your sector, your name, your competitors and whatever fragments exist. Nothing is wrong at source, because there is no source. The gap itself is the cause.

Understanding how the source ecosystem works is the difference between fixing this and redecorating your homepage.

Is it worse to be described wrongly than not described at all?

Usually, yes, though not always.

Invisibility costs you a chance to be considered. Inaccuracy costs you the consideration you already had, and it does so with the engine’s credibility attached. A buyer who does not see you might still find you another way. A buyer who is told you do not serve their sector has been given a reason to stop looking, and they will not verify it, because the answer arrived with the confident, sourced-looking framing that makes AI output feel settled.

There is also a compounding effect. A wrong claim that sits unchallenged gets picked up, restated and re-indexed elsewhere, and each restatement makes it look better corroborated. Corroboration is one of the strongest signals in retrieval, which means a false claim can accumulate exactly the property that makes engines trust it.

The exception is genuinely trivial detail. If an engine gets your founding year wrong by two years, nobody has ever lost a deal over it, and the effort is better spent elsewhere. Which is why the next question matters more than the panic does.

The FACT loop: how corrections actually work

Four steps, in order. Skipping any of them is why most attempts at this fail.

1. Find. Establish what is actually being said, at scale, rather than from one screenshot.
2. Attribute. Work out where it came from. This splits into the two root causes, and it is the step almost everyone skips.
3. Correct. Fix the wrong source, or supply the source that was missing. Different problems, different work.
4. Track. Re-measure and confirm the answer changed. A correction you cannot evidence is a correction you cannot bill, budget for, or repeat.

How do I find out what AI is actually saying about my brand?

Not by typing your company name into ChatGPT once and reading what comes back.

That tells you what one model said, in one session, to one phrasing, on one day, possibly influenced by your own chat history. It is a starting point and it is worth doing, but it is not a measurement, and treating it as one leads teams to chase errors that appear in one answer in twenty while missing errors that appear in nineteen.

If you want the manual version before anything is automated, how to check what ChatGPT says about your company walks through an hour of it across the engines.

What you need is a tracked prompt set: a fixed group of buyer-realistic questions, run repeatedly across engines, with the answers stored so you can see patterns rather than moments.

For accuracy work specifically, your prompt set needs categories that a visibility-focused set usually will not have:

• Identity prompts. “What does [company] do?” “Who are [company]?” “Is [company] still trading?”
• Capability prompts. “Does [company] work with [sector]?” “Does [company] offer [service]?”
• Comparison prompts. “How do [company] and [competitor] differ?” These surface misattribution faster than anything else.
• Credential prompts. Certifications, accreditations, memberships, locations, and anything a procurement form asks for.
• Disqualifying prompts. The questions a buyer asks when they are looking for a reason to rule you out.

Run each across the engines your buyers actually use, repeatedly, over weeks. Building a prompt set that measures something real is its own discipline, and it is worth doing properly, because everything downstream depends on the quality of what you find here.

Where did the model get it? The two root causes

This is the most important section on this page, and it is the one the published advice on this subject almost entirely misses.

Most guides tell you to correct the source. That instruction assumes a wrong source exists. Often it does not.

The tell for Cause B is drift. Ask the same question five times and read the answers side by side. A retrieved fact stays put, because it is being copied. An inferred fact wobbles, because it is being generated fresh each time from thin evidence. If the wrong claim keeps changing its details while staying wrong in the same direction, you are almost certainly looking at inference, and you can stop hunting for a page to correct.

Distinguishing these two is also why mentions and citations need reading separately. A brand can be described constantly and cited almost never, and that pattern is itself diagnostic: it usually means the engine is talking about you from a general impression rather than from a specific retrievable source.

How bad is it? Triage before you spend anything

Not every error deserves a budget. Sort by commercial consequence.

Two rules for using this well.

Weight by frequency as well as severity. A deal-blocking claim appearing in two percent of answers is a lower priority than a credibility-denting one appearing in seventy percent. Severity tells you how much a single occurrence costs. Frequency tells you how often you are paying it.

Weight by prompt, not by engine. An error in an identity or capability prompt matters more than the same error in a peripheral question, because those are the prompts sitting closest to a live buying decision. Prioritising by proximity to buyer intent applies to accuracy work exactly as it applies to page work.

How do I correct it when a wrong source exists?

Work outward from the sources with the most authority, not the ones that are easiest to reach.

Start with structured entity data. Directories, databases and profiles carry disproportionate weight because they are structured, machine-readable and treated as reference. For a UK business that usually means your Companies House record, Google Business Profile, LinkedIn company page, Crunchbase, Wikidata, and the trade body or industry directory listings for your sector. These are frequently years out of date because nobody owns them internally, and they are the cheapest high-authority corrections available to you.

Then the third-party pages carrying the claim. Trade press, partner sites, aggregators, review platforms, old press releases. Contact the publisher and ask for a correction, giving them the correct fact in a form they can paste. Publishers correct factual errors far more readily than people expect, particularly trade press. Where a page cannot be corrected, aim to supersede it: a more recent, better-corroborated page stating the correct fact will usually outweigh an older one, though not instantly.

Then your own site, last. Not because it does not matter, but because it is rarely the cause and fixing it first creates the false impression that the job is done. Its real role here is to be the unambiguous reference the corrected sources point back to, which is what citation fidelity is about: giving engines a version of the fact that is clear enough to retrieve and hard to misread.

State corrected facts plainly and in one place. “Tilio is a UK answer engine optimisation agency working with B2B brands across the UK” is retrievable. The same information spread across three paragraphs of positioning copy is not.

What do I do when there is no source to correct?

You build the source. This is slower, it is less satisfying, and it is the situation most UK B2B firms are actually in.

When no source supports the claim, publish the correct fact clearly and make related profiles consistent. That supplies useful evidence without assuming you have identified the model’s original reasoning.

Answer the question explicitly somewhere retrievable. If engines keep getting your sector focus wrong, you probably do not have a page that states your sector focus in plain, extractable language. Not implied through case studies. Stated. The facts that get inferred wrongly are almost always the facts nobody has written down flatly.

Get it corroborated off your own domain. A fact that appears only on your website is a claim. The same fact appearing on your website, your LinkedIn, a directory listing, a trade publication and a partner page is a pattern, and retrieval systems weight patterns. This is the slow part, and it is the part that actually moves the answer.

Make the entity unambiguous. Consistent naming, consistent description, consistent category language across every property you control. Inconsistency is itself a cause of inferred error: if you are described five different ways in five places, the engine will synthesise a sixth.

Accept the timescale. Building corroboration where none exists is a quarter of work, not a fortnight. Anyone promising to fix an inference problem in a week is describing the wrong problem. This is the same work as building a defensible position in zero-click search, and it compounds in the same way.

How long does it take, and who does the work?

Nobody answers this honestly, so here it is with the uncertainty left in.

Three things drive the variance and you control none of them: how often each engine recrawls the source, how heavily it weights that source against others, and whether cached or memorised material is still in play. Anyone quoting you a fixed timeline is quoting a guess.

What you can control is sequencing. Fix the structured data first, because it is fast and cheap. Start the corroboration work in parallel, because it is slow and will still be running when the quick wins have landed.

How do I prove it actually changed?

This is where most of this work quietly falls apart. The source gets fixed, everyone moves on, and nobody checks whether the answer moved.

Treat source correction and answer correction as separate observations. Record both dates. An updated source may not be retrieved promptly, and the answer may still vary between runs.

Measuring it properly means three things.

Re-run the same prompts, unchanged. If you reword the prompt between the before and after, you have measured two different things and proved nothing. The prompt set has to be fixed for the comparison to mean anything.

Measure frequency, not presence. The right question is not “does the error still appear?” but “what share of answers still contains it?” Corrections rarely go from wrong to right in one step. They go from seventy percent wrong to thirty percent wrong to occasional, and if you are only checking presence, you will read genuine progress as total failure.

Watch both the claim and the citation. Sometimes the false claim disappears while the bad source is still being cited, which means it will probably come back. Sometimes the source is dropped but the claim persists, which points to Cause B and means you have more building to do. Reading mentions and citations together is what makes the difference legible.

This is measurable, but only with the same discipline any other visibility work needs, and it is worth being clear-eyed about what platforms can and cannot actually show you. You are tracking a shifting distribution, not reading a scoreboard.

What this costs a UK B2B business

Most writing on this subject illustrates the risk with consumer examples and large-brand incidents. Those are vivid and they are not your situation.

Here is your situation. A procurement lead at a mid-sized manufacturer is building a shortlist. They ask an assistant which suppliers work with their sector. You do, extensively. The engine says you focus on something else, because a 2022 directory listing said so and nothing since has said otherwise clearly enough. You are not on the shortlist. You never find out you were considered, because there was no enquiry to lose.

That is the shape of this problem in B2B: invisible, unattributed, and absent from every dashboard you own. It does not appear as a drop in enquiries, because the enquiries were never made.

The arithmetic is uncomfortable and simple. A firm closing forty deals a year at £40,000 average contract value does not need many mis-shortlisted opportunities before the cost of the error exceeds the cost of fixing it by an order of magnitude. And unlike a paid campaign, the loss compounds, because each quarter the wrong claim survives is another quarter it gets restated somewhere new.

This is also why accuracy work usually outranks pure visibility work in priority. Being harder to find is a growth problem. Being confidently described as something you are not is a conversion problem happening before you knew a conversation had started.

Common mistakes

Fixing the website first and stopping there. The most common error by a wide margin. Your site is rarely the cause, and fixing it feels productive enough to end the project.

Treating one screenshot as evidence. It tells you something happened once. Deciding what to fix needs frequency, and frequency needs repeated tracked prompts.

Chasing every error equally. Without triage, teams spend a quarter fixing a founding year while a deal-blocking claim runs unchallenged.

Assuming a wrong source exists. If you go hunting for a page to correct when the real cause is inference, you will find nothing, conclude the problem is unfixable, and abandon work that would have succeeded with a different method.

Expecting a fast answer change. Sources update on their own schedule. Weeks is normal. Panic at week two causes teams to abandon fixes that were about to land.

Measuring with a moving prompt. Rewording the prompt between checks makes the comparison meaningless, and it is the easiest way to convince yourself something worked when it did not.

Where to start

If the error is deal-blocking, capture it properly, check whether the claim sits on a cited page, and fix the structured entity sources this week. Those are the fastest corrections available and they cost almost nothing.

If it is broader than one claim, or you suspect there is no source to correct, start with diagnosis rather than repair: what the engines are saying about you across a tracked prompt set, which claims are wrong and how often, where each one came from, and which of the two root causes you are dealing with. In Gain, Tilio’s monthly AEO programme, that diagnosis is part of onboarding.

You can add ongoing brand accuracy to your programme. We investigate inaccurate claims, work on the sources behind them and keep answers under review, within the agreed scope. See how the work is planned and reviewed.

Book a call if you would rather talk it through first.

Tilio is a UK answer engine optimisation agency. We measure what AI engines say about B2B brands, find what they get wrong, and fix it at source.